LLM-Compiled Wiki

A personal knowledge base that an LLM maintains for you. Drop raw data in (call transcripts, emails, docs), run the compiler, and get a structured, interlinked wiki you can query instantly.

Inspired by Andrej Karpathy's approach to using LLMs as knowledge compilers.

How It Works

You drop raw files in ──> LLM compiles them ──> Wiki updates automatically
     (transcripts,          (reads new data,       (structured articles
      emails, docs)          updates articles)       you can query)

Two layers:

• raw/ - Your source material. Transcripts, emails, docs. Never edited by the LLM.
• wiki/ - LLM-compiled articles. Structured, interlinked, searchable. The LLM maintains this - you read it.

Three operations:

1. Ingest - Drop a new file in raw/, run the compiler, wiki updates
2. Query - Ask a question, LLM reads the wiki and gives you a sourced answer
3. Lint - Periodic health check finds contradictions, stale content, broken links

Quick Start

1. Set up your project

Open this folder in Claude Code or as a Cowork project. The CLAUDE.md file tells Claude how to use the wiki.

2. Drop your first raw files

Put some source material in the right folder:

• Call transcripts/meeting notes -> raw/calls/
• Email summaries -> raw/email/
• Everything else (docs, specs, research) -> raw/docs/

Name files: YYYY-MM-DD_Description_Type.ext (e.g., 2025-06-15_Sprint-Review_Summary.md)

3. Run the first compile

Tell Claude: "compile the wiki"

It will read your raw files, generate wiki articles, and populate the indexes. Your first compile builds the wiki from scratch. After that, compiles are incremental - only new data gets processed.

4. Start querying

Ask questions like:

• "What's the current state of Project X?"
• "What decisions have we made about integrations?"
• "What has Sarah been involved in this quarter?"
• "What's on my plate this week?"

Claude reads the wiki index, finds the relevant articles, and gives you a sourced answer.

5. Keep feeding it

The more raw data you add, the smarter the wiki gets. Set up automated ingest for email (scheduled task) and transcripts (file watcher or manual drop).

Directory Structure

llm-wiki/
├── CLAUDE.md                  <- Instructions for Claude (routing, rules, conventions)
├── plan.md                    <- Implementation guide and phasing
├── build_wiki.py              <- Static HTML site generator (optional)
├── skills/
│   ├── compile-wiki/SKILL.md  <- How to compile raw data into wiki articles
│   └── lint-wiki/SKILL.md     <- How to run wiki health checks
├── wiki/
│   ├── INDEX.md               <- Top-level index (section pointers only)
│   ├── log.md                 <- Append-only changelog
│   ├── projects/              <- One article per project
│   ├── people/                <- One article per person
│   ├── decisions/             <- Major decisions, date-prefixed
│   ├── concepts/              <- Cross-cutting patterns and ideas
│   └── timeline/              <- Quarterly summaries
├── voice/                     <- Talk to your wiki (realtime voice, see voice/README.md)
└── raw/
    ├── .manifest.json         <- Tracks what's been compiled
    ├── calls/                 <- Meeting transcripts, summaries
    ├── email/                 <- Email digests
    ├── docs/                  <- Specs, research, strategy docs
    └── slack/                 <- Channel summaries (future)

Wiki Article Types

Projects

Track initiatives with status, ownership, decisions, open questions, timeline, and source links.

People

Cross-project profiles showing what each person owns, their recurring positions, and recent activity.

Decisions

Date-stamped records of significant choices - what was decided, why, and what it affected.

Concepts

Patterns and ideas that span multiple projects (e.g., "governed content creation", "feature flags").

Timeline

Quarterly summaries that capture what happened, what was planned, and how reality diverged from the plan.

Key Principles

1. Raw stays raw. Never modify source material.
2. The LLM is the author. You direct it, but don't hand-edit wiki articles.
3. Incremental compilation. New data updates existing articles - no full rewrites.
4. Everything backlinks. Articles reference each other with [[slug]] links.
5. TLDRs save tokens. Every article has a one-line TLDR so Claude can decide whether to read the full article.
6. Index files are lightweight. Top-level index is ~10 lines. Full listings live in section indexes.
7. Log everything. The changelog (log.md) tracks every compile and structural change.

Optional: Static Site Generator

build_wiki.py converts the wiki into a single self-contained HTML file you can open in a browser or share on Slack. It handles markdown conversion, backlink resolution, search indexing, and privacy filtering.

pip install markdown
python build_wiki.py
open wiki-site.html

Optional: Talk to Your Wiki 🎙️

voice/ adds a realtime voice interface: press a global hotkey, ask a question out loud, and hear your wiki answer back. It connects to OpenAI's gpt-realtime-2 speech-to-speech model and grounds every answer in wiki/ via search_wiki / read_article tools — the same INDEX → TLDR → article read-path Claude uses.

cd voice
./setup.sh                 # installs everything, creates .env
# add your OpenAI API key to voice/.env, then:
./run.sh                   # press Ctrl+Option+W to talk

Or hand the repo to Claude and say "set up and run the wiki voice, my OpenAI key is sk-..." — it'll do the rest. See voice/README.md for permissions and options.

Customization

Adapt the sections

The five sections (projects, people, decisions, concepts, timeline) work for product/engineering teams. Adapt them to your domain:

• Research team: papers, authors, methods, findings, timeline
• Sales team: deals, contacts, competitors, playbooks, timeline
• Personal: goals, people, decisions, interests, journal

Add data sources

The system handles any text-based source. Common additions:

• Slack channel summaries (daily digest -> raw/slack/)
• Jira/Linear exports
• CRM notes
• Interview transcripts

Privacy filtering

The build script supports stripping sensitive content (e.g., 1:1 quotes) when generating the HTML site. Customize the filter patterns in build_wiki.py.

Works With

• Claude Code (terminal) - best for compiling, linting, structural changes
• Claude Cowork (desktop) - best for ad-hoc queries, single-file ingests, daily briefings
• Obsidian - great for browsing the wiki visually (backlinks, graph view)
• Any markdown viewer - the wiki is just markdown files